Predictive Maintenance with AI
Artificial intelligence (AI) still seems like science fiction, with talking machines and sentient robots—books and films have been shaping this image for years. But AI has long since arrived in the present, with practical areas of application.
Where large amounts of data and complex interrelationships come together, computer-aided tools and solutions help to evaluate and organize them. In the industrial environment, these are typically digital production control systems, officially known as Manufacturing Execution Systems (MES). Such systems ensure data collection and evaluation, creating a digital image of production. This in turn enables effective control, management, and analysis of holistic manufacturing processes and opens up potential for optimization. This is based on large amounts of data collected from machines and workstations, in warehousing and transport processes, and much more. MES solutions work very effectively and help companies to improve processes in the long term – artificial intelligence offers additional, extremely high growth potential here, as it enables predictive and autonomous control processes.
The following interview was conducted in March 2021 as part of the collaboration between aiXbrain GmbH and GFOS mbH. Our aiXbrain managing director Alexander Engels was asked about the possibilities of AI for maintenance.
Predictive maintenance with the help of artificial intelligence
If malfunctions or failures occur in machines or production lines, this is also visible in the recorded data, because machine and measurement values deviate from target values and this data flows into the databases. MES tools analyze the data so that possible sources of error can be identified and rectified. But what if it were possible to issue a warning or even make a correction at an early stage, before a production failure threatens to occur?
Dr. Engels, first of all, thank you very much for taking the time to do this interview. Before we get to the above question, let's start at the beginning: Can you give us a rough timeline and technological overview of artificial intelligence and tell us why AI solutions are becoming so widespread right now?
Dr. Alexander Engels: The origins of artificial intelligence go back a long way. Ideas about artificial neural networks, for example, emerged as early as the middle of the last century. However, their practical implementation and meaningful application were closely linked to the availability of data and computing power from the outset. It is therefore not surprising that the use of neural networks, for example for language processing, was still rather ridiculed even at the turn of the millennium. Just a few years later, however, the mass use of the internet, computing power in the cloud and on graphics cards, and the development of some clever methods for training neural networks led to a real breakthrough in automated speech and image processing. Since then, further models and methods of machine learning have been developed and the focus of application has expanded to almost all areas of work and life. Now we have begun to network everything and make machine data available for AI. The starting signal for this was not too long ago, perhaps one or two years. However, the speed of the race for the value-adding use of AI in everyday production will soon eclipse everything that has gone before.
Are there different forms of artificial intelligence, and can this be illustrated with application scenarios?
Dr. Alexander Engels: Since artificial intelligence is basically a composite technology consisting of mathematics, computer science, computing hardware, and, in some cases, application fields, there are a whole range of different definitions, from technical to philosophical. Personally, I would distinguish AI according to its purpose, namely in three basic disciplines: First, pattern recognition, for example to detect irregularities or diagnose malfunctions. Second, forecasting, for example for the early detection of impending material defects or problems in the process flow. And thirdly, control, i.e., the submission of concrete proposals for action by an AI, for example, for parameterizing planning processes or controlling machines. A typical and thoroughly sensible AI path for companies along these basic disciplines would probably be "recognize, understand, predict, counteract."
Put simply, AI systems use data to find patterns and regularities and learn from them. If a data basis already exists, as is the case with current MES solutions, for example, the AI tool can learn very effectively. But how should a layperson imagine this learning process?
Dr. Alexander Engels: Basically, it's not much different from classic human learning in everyday life or during training. The AI apprentice goes to AI school, so to speak. The only difference is that the AI apprentice can learn many times faster, provided it finds enough illustrative material of sufficient quality. Like its human counterpart, the AI apprentice learns from the things it observes, or rather from the data and information it is fed during its training. Or to put it plainly: the apprentice does not know anything it has never seen in any form. With each new data set, the AI apprentice calibrates its internal adjustment screws, i.e., the parameters of the underlying AI model. When the adjustment screws change only slightly, the apprenticeship is complete. Readjustments are then only made when new things happen, but this is already part of everyday work. However, the AI apprentice lacks the mental transfer capabilities that a human being can perform once they have reached a certain level of knowledge. It is therefore focused exclusively on a clearly defined area of responsibility. Without going into detail here, it should be noted that there is a particularly exciting area of AI research known as transfer learning.
Artificial intelligence can generally be used in any area that works with data. What do you find particularly appealing about its application in the context of the manufacturing industry?
Dr. Alexander Engels: With its highly complex human-machine systems and processes that are sometimes timed down to the second, the manufacturing industry is incredibly diverse and challenging. Every factory has its own life; no two workshops are alike in terms of equipment and processes. Nevertheless, over the years and decades, clever people and skilled engineers have managed to set up even the most complex factories to operate efficiently and effectively. However, the price for this is often a certain rigidity in the operating procedures, where everything is precisely coordinated. Just like the famous Swiss watch movement. But if things such as machine availability, material supply, or cycle times deviate too far from the target value in daily operations, the planned processes no longer function and there are sometimes massive cross-effects on the entire movement. And if people are unable to immediately correct these cross-effects in such situations, money is lost. The manufacturing industry is therefore characterized by its high complexity, large amounts of data, and great potential for value creation. This is the perfect playground for artificial intelligence, which is why we are getting involved.
We started with the example of AI-based maintenance that detects impending malfunctions or failures at an early stage. This allows maintenance and repair measures to be carried out in good time, ensuring that production runs smoothly. Can you describe and explain this scenario in a little more detail?
Dr. Alexander Engels: I already described the initial scenario in the factory in the previous point, i.e., the perfectly planned Swiss clockwork. Now let's imagine that the clockwork is divided into several consecutive production stages. These stages must mesh with each other, one after the other, so that the raw material at the factory entrance becomes the end product at the factory exit. And at this very exit, the customer is already waiting and pushing, because, as we know, they should have received their delivery yesterday. In this very situation, a machine unexpectedly breaks down in the penultimate production step. And, of course, it is precisely the machine on which the product the customer is waiting for is running. And, of course, this all happens during the night shift, when neither the planning department nor the staff are there to quickly convert an alternative machine. Wouldn't it be wonderful if there were an "oracle" that would have pointed out a possible problem in the night shift during the morning shift and perhaps even provided a diagnosis of exactly what was likely to go wrong? A well-trained maintenance AI has precisely these capabilities and also conveniently works as a digital assistant in the MES.
In your opinion, how important is tool management for a smooth production process? Is knowledge about wear and tear and impending failure particularly valuable in this context?
Dr. Alexander Engels: Especially in manufacturing processes such as cold forming or cutting, tools are usually used for "steel on steel" applications. This means that the service life of such tools is limited. So the question is always: how long or how many workpieces will the tool last before it needs to be replaced? In the case of long-running processes or large batches, it always involves a certain risk for production planners to plan upcoming shifts with tools that have already been in use for a long time. On the other hand, tools or tool parts are often a scarce commodity, as they cost money and tie up capital. The tricky thing about tool wear is that it effectively depends on a whole range of factors, such as raw material, stroke, force design, number of start-up processes, etc. Mere statistics such as mean values are therefore insufficient and unreliable. An AI forecast of effective tool wear that takes all available peripheral information into account can shed light on the situation. It may not be the best news when the planner learns in the morning that important tools will not be available throughout the day. But that's always better than a breakdown during the night shift and chaos the next morning.
Maintenance and the large area of tool management are areas that generally offer a lot of potential for optimization, as malfunctions and resource shortages always mean production downtime and thus a loss of money. What other areas and processes within production do you consider to be extremely suitable for AI?
Dr. Alexander Engels: Basically, AI offers a great opportunity to solve all the small and large problems in operational processes that every manufacturing company is familiar with and that previously could only be dealt with at great expense, using a completely new toolbox. The prerequisite for this is sufficient data and the willingness of the user to embrace the new and perhaps somewhat counterintuitive AI assistants. If both of these conditions are met, we are talking about the use of AI for quality assurance, directly in the manufacturing process at the machine and at the earliest possible stages of the value chain. Then we are talking about perfecting material flow planning, i.e., ensuring the cost-efficient availability of raw materials, semi-finished products, and personnel. Or even things like identifying and eliminating bottlenecks in material purchasing or inventory management long before they actually occur. It's easy to get carried away and perhaps lose yourself in thoughts that are too far removed from the status quo. It is important to start at a sensible point, use the experience gained to design and further develop an AI roadmap, get employees on board with the process, and, of course, bring competent and reliable AI partners on board.
Dr. Engels, thank you very much for talking to us – we are sure that there will be more exciting interviews to follow, shedding even more light on artificial intelligence and its possibilities for industry.
A powerful partnership for your success
aiXbrain GmbH has entered into a close cooperation with GFOS mbH to supplement its proven IT solutions in the field of manufacturing execution systems with professional AI modules. The decisive factors for the cooperation were the high level of technical expertise and the pronounced willingness to innovate on both sides – this will enable manufacturing companies to benefit from powerful software solutions in the long term and optimize their production processes sustainably.
You can find out more about AI and our partnership with GFOS in an interview with our CTO Simon Görtzen and Mr. Röhrig, Chairman of the GFOS Advisory Board.
